trsm_kernel_lt_4x4_barcelona.s

Optimized GotoBLAS libraries

	movddup	 -6 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm4
	movddup	 -7 * SIZE(BO), %xmm9
	mulpd	 %xmm4, %xmm9
	subpd	 %xmm9, %xmm2
	movddup	 -8 * SIZE(BO), %xmm10
	mulpd	 %xmm4, %xmm10
	subpd	 %xmm10, %xmm0

	movddup	-11 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm2
	movddup	-12 * SIZE(BO), %xmm9
	mulpd	 %xmm2, %xmm9
	subpd	 %xmm9, %xmm0

	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm0
#endif

#ifdef LN
	subq	$2 * SIZE, CO1
	subq	$2 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movlpd	%xmm9,   0 * SIZE(CO1)
	movlpd	%xmm13,  1 * SIZE(CO1)

	movhpd	%xmm9,   0 * SIZE(CO2)
	movhpd	%xmm13,  1 * SIZE(CO2)

	movlpd	%xmm11,  0 * SIZE(CO1, LDC, 2)
	movlpd	%xmm15,  1 * SIZE(CO1, LDC, 2)

	movhpd	%xmm11,  0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm15,  1 * SIZE(CO2, LDC, 2)
#else
	movlpd	%xmm0,  0 * SIZE(CO1)
	movhpd	%xmm0,  1 * SIZE(CO1)

	movlpd	%xmm2,  0 * SIZE(CO2)
	movhpd	%xmm2,  1 * SIZE(CO2)

	movlpd	%xmm4,  0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm4,  1 * SIZE(CO1, LDC, 2)

	movlpd	%xmm6,  0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm6,  1 * SIZE(CO2, LDC, 2)
#endif

#if defined(LN) || defined(LT)
	movaps	%xmm9,  -16 * SIZE(BO)
	movaps	%xmm11, -14 * SIZE(BO)
	movaps	%xmm13, -12 * SIZE(BO)
	movaps	%xmm15, -10 * SIZE(BO)
#else
	movaps	%xmm0,  -16 * SIZE(AO)
	movaps	%xmm2,  -14 * SIZE(AO)
	movaps	%xmm4,  -12 * SIZE(AO)
	movaps	%xmm6,  -10 * SIZE(AO)
#endif

#ifndef LN
	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 4), BO
#endif

#ifdef LN
	subq	$2, KK
#endif

#ifdef LT
	addq	$2, KK
#endif

#ifdef RT
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif
	ALIGN_4	

.L30:
	testq	$1, M
	je	.L39

#ifdef LN
       movq	K, %rax
       salq	$0 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	movq	AORIG, AO
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
#endif

	movq	B, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(BO, %rax, 4), BO
#endif	

	movddup	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movddup	-14 * SIZE(AO), %xmm2
	pxor	%xmm9, %xmm9
	movddup	-15 * SIZE(AO), %xmm4
	pxor	%xmm10, %xmm10
	movapd	-16 * SIZE(BO), %xmm1
	pxor	%xmm11, %xmm11
	movapd	 -8 * SIZE(BO), %xmm3

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
	negq	%rax
	je,pn	.L36
	ALIGN_4

.L32:
	mulpd	%xmm0, %xmm1
	mulpd	-14 * SIZE(BO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm9
	movddup	-12 * SIZE(AO, %rax, 1), %xmm0
	mulpd	%xmm4, %xmm1
	mulpd	-10 * SIZE(BO, %rax, 4), %xmm4
	addpd	%xmm1, %xmm10
	movapd	  (BO, %rax, 4), %xmm1
	addpd	%xmm4, %xmm11
	movddup	-11 * SIZE(AO, %rax, 1), %xmm4
	mulpd	%xmm2, %xmm3
	mulpd	 -6 * SIZE(BO, %rax, 4), %xmm2
	addpd	%xmm3, %xmm8
	movapd	 -4 * SIZE(BO, %rax, 4), %xmm3
	addpd	%xmm2, %xmm9
	movddup	-13 * SIZE(AO, %rax, 1), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	 -2 * SIZE(BO, %rax, 4), %xmm2
	addpd	%xmm3, %xmm10
	movapd	  8 * SIZE(BO, %rax, 4), %xmm3
	addpd	%xmm2, %xmm11
	movddup	-10 * SIZE(AO, %rax, 1), %xmm2

	addq	$4 * SIZE, %rax
	jl,pt	.L32
	ALIGN_4

.L36:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L38

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
	negq	%rax
	ALIGN_4

.L37:
	mulpd	%xmm0, %xmm1
	mulpd	-14 * SIZE(BO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm9
	movddup	-15 * SIZE(AO, %rax, 1), %xmm0

	addq	$SIZE, %rax
	jl	.L37
	ALIGN_4

.L38:
	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9

#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$1, %rax
#else
	subq	$4, %rax
#endif

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 4), BO
#endif

#if defined(LN) || defined(LT)
	movapd	-16 * SIZE(BO), %xmm2
	movapd	-14 * SIZE(BO), %xmm3

	subpd	%xmm8,  %xmm2
	subpd	%xmm9,  %xmm3
#else
	movapd	-16 * SIZE(AO), %xmm2
	movapd	-14 * SIZE(AO), %xmm3

	subpd	%xmm8, %xmm2
	subpd	%xmm9, %xmm3
#endif

#if defined(LN) || defined(LT)
	movddup	-16 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm2
	mulpd	 %xmm0, %xmm3
#endif

#ifdef RN
	movapd	%xmm2, %xmm0
        unpckhpd %xmm0, %xmm0

	movapd	%xmm3, %xmm1
        unpckhpd %xmm1, %xmm1

	movsd	-16 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm2

	movsd	-15 * SIZE(BO), %xmm5
	mulsd	 %xmm2, %xmm5
	subsd	 %xmm5, %xmm0
	movsd	-14 * SIZE(BO), %xmm6
	mulsd	 %xmm2, %xmm6
	subsd	 %xmm6, %xmm3
	movsd	-13 * SIZE(BO), %xmm7
	mulsd	 %xmm2, %xmm7
	subsd	 %xmm7, %xmm1

	movsd	-11 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm0

	movsd	-10 * SIZE(BO), %xmm5
	mulsd	 %xmm0, %xmm5
	subsd	 %xmm5, %xmm3
	movsd	 -9 * SIZE(BO), %xmm6
	mulsd	 %xmm0, %xmm6
	subsd	 %xmm6, %xmm1

	movsd	 -6 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm3

	movsd	 -5 * SIZE(BO), %xmm5
	mulsd	 %xmm3, %xmm5
	subsd	 %xmm5, %xmm1

	movsd	 -1 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm1

	unpcklpd %xmm0, %xmm2
	unpcklpd %xmm1, %xmm3
#endif

#ifdef RT
	movapd	%xmm2, %xmm0
        unpckhpd %xmm0, %xmm0

	movapd	%xmm3, %xmm1
        unpckhpd %xmm1, %xmm1

	movsd	 -1 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm1

	movsd	 -2 * SIZE(BO), %xmm5
	mulsd	 %xmm1, %xmm5
	subsd	 %xmm5, %xmm3
	movsd	 -3 * SIZE(BO), %xmm6
	mulsd	 %xmm1, %xmm6
	subsd	 %xmm6, %xmm0
	movsd	 -4 * SIZE(BO), %xmm7
	mulsd	 %xmm1, %xmm7
	subsd	 %xmm7, %xmm2

	movsd	 -6 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm3

	movsd	 -7 * SIZE(BO), %xmm5
	mulsd	 %xmm3, %xmm5
	subsd	 %xmm5, %xmm0
	movsd	 -8 * SIZE(BO), %xmm6
	mulsd	 %xmm3, %xmm6
	subsd	 %xmm6, %xmm2

	movsd	-11 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm0

	movsd	-12 * SIZE(BO), %xmm5
	mulsd	 %xmm0, %xmm5
	subsd	 %xmm5, %xmm2

	movsd	-16 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm2

	unpcklpd %xmm0, %xmm2
	unpcklpd %xmm1, %xmm3

#endif

#ifdef LN
	subq	$1 * SIZE, CO1
	subq	$1 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movlpd	%xmm2,  0 * SIZE(CO1)
	movhpd	%xmm2,  0 * SIZE(CO2)
	movlpd	%xmm3,  0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm3,  0 * SIZE(CO2, LDC, 2)
#else
	movlpd	%xmm2,  0 * SIZE(CO1)
	movhpd	%xmm2,  0 * SIZE(CO2)
	movlpd	%xmm3,  0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm3,  0 * SIZE(CO2, LDC, 2)
#endif

#if defined(LN) || defined(LT)
	movaps	%xmm2, -16 * SIZE(BO)
	movaps	%xmm3, -14 * SIZE(BO)
#else
	movaps	%xmm2, -16 * SIZE(AO)
	movaps	%xmm3, -14 * SIZE(AO)
#endif

#ifndef LN
	addq	$1 * SIZE, CO1
	addq	$1 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
#endif

#ifdef LN
	subq	$1, KK
#endif

#ifdef LT
	addq	$1, KK
#endif

#ifdef RT
       movq	K, %rax
       salq	$0 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif
	ALIGN_4	
	
.L39:
#ifdef LN
       leaq	(, K, SIZE), %rax
       leaq	(B, %rax, 4), B
#endif

#if defined(LT) || defined(RN)
	movq	BO, B
#endif

#ifdef RN
	addq	$4, KK
#endif

#ifdef RT
	subq	$4, KK
#endif

	decq	J			# j --
	jg	.L01
	ALIGN_4

.L40:
	testq	$2, N
	je	.L80

#if defined(LT) || defined(RN)
	movq	A, AO
#else
	movq	A, AORIG
#endif

#ifdef RT
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       subq	%rax, B

       leaq	(, LDC, 2), %rax
       subq	%rax, C
#endif

	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
#ifndef RT
	leaq	(C, LDC, 2), C
#endif

#ifdef LN
	movq	OFFSET, %rax
	addq	M, %rax
	movq	%rax, KK
#endif	

#if defined(LT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L60
	ALIGN_4

.L51:
#ifdef LN
       movq	K, %rax
       salq	$2 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	movq	AORIG, AO
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
#endif

	movq	B, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(BO, %rax, 2), BO
#endif	

	movddup	-16 * SIZE(BO), %xmm1
	movddup	-15 * SIZE(BO), %xmm5
	pxor	%xmm8, %xmm8
	movddup	-12 * SIZE(BO), %xmm3
	pxor	%xmm9, %xmm9
	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm12, %xmm12
	movapd	 -8 * SIZE(AO), %xmm4
	pxor	%xmm13, %xmm13

#ifndef LN
	prefetchw      3 * SIZE(CO1)
	movapd	%xmm0, %xmm2
	prefetchw      3 * SIZE(CO2)
#else
	prefetchw     -8 * SIZE(CO1)
	movapd	%xmm0, %xmm2
	prefetchw     -8 * SIZE(CO2)
#endif


#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	je,pn	.L56
	ALIGN_4

.L52:
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	-14 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-13 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm0
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	  (AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	 -8 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-11 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm4, %xmm2
	mulpd	%xmm3, %xmm4
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm4, %xmm8
	movapd	 -4 * SIZE(AO, %rax, 4), %xmm4
	addpd	%xmm3, %xmm12
	movddup	-10 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm5, %xmm2
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	 -9 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm4, %xmm2
	mulpd	%xmm3, %xmm4
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm4, %xmm8
	movapd	  8 * SIZE(AO, %rax, 4), %xmm4
	addpd	%xmm3, %xmm12
	movddup	 -4 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm5, %xmm2
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	 -7 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2

	addq	$4 * SIZE, %rax
	jl,pt	.L52
	ALIGN_4

.L56:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L59

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	ALIGN_4

.L57:
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	-14 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-13 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2

	addq	$SIZE, %rax
	jl	.L57
	ALIGN_4

.L59:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$4, %rax
#else
	subq	$2, %rax
#endif

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 2), BO
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm8, %xmm0
	unpcklpd %xmm9, %xmm8
	unpckhpd %xmm9, %xmm0

	movapd	%xmm12, %xmm4
	unpcklpd %xmm13, %xmm12
	unpckhpd %xmm13, %xmm4

	movapd	-16 * SIZE(BO), %xmm9
	movapd	-14 * SIZE(BO), %xmm13
	movapd	-12 * SIZE(BO), %xmm1
	movapd	-10 * SIZE(BO), %xmm5

	subpd	%xmm8,  %xmm9
	subpd	%xmm0,  %xmm13
	subpd	%xmm12, %xmm1
	subpd	%xmm4,  %xmm5
#else
	movapd	-16 * SIZE(AO), %xmm0
	movapd	-14 * SIZE(AO), %xmm1
	movapd	-12 * SIZE(AO), %xmm2
	movapd	-10 * SIZE(AO), %xmm3

	subpd	%xmm8, %xmm0
	subpd	%xmm12, %xmm1
	subpd	%xmm9, %xmm2
	subpd	%xmm13, %xmm3
#endif

#ifdef LN
	movddup	 -1 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm5
	movddup	 -2 * SIZE(AO), %xmm10
	mulpd	 %xmm5, %xmm10
	subpd	 %xmm10, %xmm1
	movddup	 -3 * SIZE(AO), %xmm12
	mulpd	 %xmm5, %xmm12
	subpd	 %xmm12, %xmm13
	movddup	 -4 * SIZE(AO), %xmm14
	mulpd	 %xmm5, %xmm14
	subpd	 %xmm14, %xmm9

	movddup	 -6 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm1
	movddup	 -7 * SIZE(AO), %xmm10
	mulpd	 %xmm1, %xmm10
	subpd	 %xmm10, %xmm13
	movddup	 -8 * SIZE(AO), %xmm12
	mulpd	 %xmm1, %xmm12
	subpd	 %xmm12, %xmm9

	movddup	-11 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm13
	movddup	-12 * SIZE(AO), %xmm10
	mulpd	 %xmm13, %xmm10
	subpd	 %xmm10, %xmm9

	movddup	-16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9
#endif

#ifdef LT
	movddup -16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9
	movddup	-15 * SIZE(AO), %xmm10
	mulpd	 %xmm9, %xmm10
	subpd	 %xmm10, %xmm13
	movddup	-14 * SIZE(AO), %xmm12
	mulpd	 %xmm9, %xmm12
	subpd	 %xmm12, %xmm1